---
license: apache-2.0
task_categories:
- text-to-image
language:
- ar
tags:
- arabic
- Qari
- OCR
- ArabicOCR
- BookStyle
- Markdown
pretty_name: Qari-OCR
size_categories:
- 10K<n<100K
---

# QARI Markdown Mixed Dataset

<div align="center">

![QARI OCR](https://img.shields.io/badge/QARI-OCR-blue)
![Arabic](https://img.shields.io/badge/Language-Arabic-green)
![Dataset](https://img.shields.io/badge/Type-Synthetic_Dataset-orange)
![License](https://img.shields.io/badge/License-Apache_2.0-red)

</div>

<div align="center">
 <img src="https://cdn-uploads.huggingface.co/production/uploads/628f7a71dd993507cfcbe587/Txz_HjVy6NsdmcghXqVH_.png" alt="QARI Logo" width="400">
</div>

## 📋 Dataset Summary

The QARI v0.3 Markdown Mixed Dataset is a specialized synthetic dataset designed for training Arabic OCR models with a focus on complex document layouts and HTML structure understanding. 
This dataset is part of the QARI-OCR project, which achieves state-of-the-art performance in Arabic text recognition.

This dataset contains **37,000 synthetically generated Arabic document images** (29.6k train, 3.7k validation, 3.7k test) with corresponding ground truth text in HTML/Markdown format, featuring:
- 🔤 **Full diacritical marks (tashkeel)** support
- 📝 **Mixed font sizes** within documents (headers, body text, annotations)
- 🎨 **12 distinct Arabic fonts** ranging from common Naskh to ornate calligraphic styles
- 📄 **Realistic document layouts** with structural HTML tags
- 🖼️ **Multiple text sources** including Basma2423 and YoussefAnwar Arabic news

## 🎯 Intended Use

This dataset is specifically designed for:
- Training OCR models that need to understand document structure
- Fine-tuning vision-language models for Arabic text recognition
- Developing systems that preserve formatting and layout information
- Research in Arabic document analysis and understanding

## 📊 Dataset Statistics

| Metric | Value |
|--------|-------|
| **Total Images** | 37,000 |
| **Train Set** | 29,600 (80%) |
| **Validation Set** | 3,700 (10%) |
| **Test Set** | 3,700 (10%) |
| **Text Sources** | oddadmix/Basma2423-Text-with-Diacritics-Correction + YoussefAnwar/Arabic-news |
| **Font Variety** | 12 Arabic fonts |
| **Font Size Range** | 14px - 100px |
| **Diacritics Support** | ✅ Full tashkeel |
| **HTML Structure** | ✅ Preserved |
| **Layout Complexity** | ✅ High (mixed sizes, headers) |

## 🔧 Data Generation Pipeline

<div align="center">

| Stage | Process | Details |
|-------|---------|---------|
| **1. Text Collection** | Source gathering | Basma2423 (with diacritics) + YoussefAnwar Arabic news |
| **2. HTML Templating** | Layout generation | Mixed font sizes, structural elements |
| **3. Rendering** | WeasyPrint → PDF → Image | High-quality document rendering |
| **4. Degradation** | Synthetic noise | Clean / Moderate / Heavy variants |

</div>

## 📈 Model Performance

When used to train QARI v0.3, this dataset enables:

| Metric | Score |
|--------|-------|
| **Character Error Rate (CER)** | 0.300 |
| **Word Error Rate (WER)** | 0.485 |
| **BLEU Score** | 0.545 |
| **Training Time** | 11 hours |
| **CO₂ Emissions** | 1.88 kg eq. |

### Key Advantages:
- 📐 **Superior layout understanding** compared to plain text models
- 🏷️ **HTML tag preservation** for structured document conversion
- ⚡ **Resource efficient** - 5x less training time than larger datasets
- 🎯 **Specialized performance** for document structure tasks

## Citation

```markdown
@article{wasfy2025qari,
  title={QARI-OCR: High-Fidelity Arabic Text Recognition through Multimodal Large Language Model Adaptation},
  author={Wasfy, Ahmed and Nacar, Omer and Elkhateb, Abdelakreem and Reda, Mahmoud and Elshehy, Omar and Ammar, Adel and Boulila, Wadii},
  journal={arXiv preprint arXiv:2506.02295},
  year={2025}
}
```